Protein backbone and sidechain torsion angles predicted from NMR chemical shifts using artificial neural networks

Abstract

A new program, TALOS-N, is introduced for predicting protein backbone torsion angles from NMR chemical shifts. The program relies far more extensively on the use of trained artificial neural networks than its predecessor, TALOS+. Validation on an independent set of proteins indicates that backbone torsion angles can be predicted for a larger, ≥90 % fraction of the residues, with an error rate smaller than ca 3.5 %, using an acceptance criterion that is nearly two-fold tighter than that used previously, and a root mean square difference between predicted and crystallographically observed (ϕ, ψ) torsion angles of ca 12º. TALOS-N also reports sidechain χ(1) rotameric states for about 50 % of the residues, and a consistency with reference structures of 89 %. The program includes a neural network trained to identify secondary structure from residue sequence and chemical shifts.

Fig. 1

Average (φ,ψ)-dependence of the Δδ¹³C^α chemical shift in Thr for three different χ¹ rotameric states: (a) g+; (b) g−; and (c) t, displayed as Ramachandran maps. Only regions with a residue density (for definition see (Spera and Bax 1991)) larger than one are color coded. The residue density is marked by gray contour lines, increasing from 1 with an increment factor of 3.2. The average secondary chemical shifts for residues in the α regions are 0.6±2.0, 3.9±1.7 and 2.0±−2.4 ppm, for χ¹ rotameric states of g+, g− and t, respectively, and −2.2±1.2, −0.2±1.4, and −2.2±1.5 ppm, respectively, for residues in the β region.

Fig. 2

(φ,ψ)-ANN predicted (φ,ψ) likelihood distributions presented as Ramachandran maps for residues 7 to 10 of protein GB3. Only 20°×20° voxels with predicted likelihoods that fall at least one standard deviation above the average population (1/324) are color coded. The φ/ψ angles observed in the reference structure (first conformer of PDB entry 2OED) is marked with a white circle; the φ/ψ angles of the center residue of the 25 best matched database fragments are displayed as green dots. The horizontal axis of each plot corresponds to φ (ranging from −180° to 180°) with the vertical axis being ψ (ranging from −180° to 180°, bottom to top). Residue G9 shows two clusters; one centered near (φ,ψ) = (−100°, 180°) and one at (120°, 160°).

Fig. 3

Flow diagram for the TALOS-N program, with the left branch corresponding to prediction of backbone torsion angles, and the right branch dedicated to χ¹.

Fig. 4

TALOS-N graphic user interface, displaying results for residue L8 of query protein ubiquitin. The left panel shows a plot of the φ/ψ angles of the 25 closest database matches (green symbols), superimposed on a Ramachandran map depicting in gray the standard most favorable backbone torsion angles for Leu. The 324 (φ,ψ)-ANN predicted scores for L8 are shown as colored voxels, but only for those that are at least one standard deviation above the average predicted voxel density. The top right panel is identical to that of the TALOS+ graphic user interface, and displays the sequence of the protein with residues marked according to their (φ,ψ) prediction classification, i.e, no prediction in light grey, consistent predictions in light or dark green (for “Strong” and “Generous” predictions, respectively), ambiguous predictions in yellow, and dynamic residues in blue. Three other panels correspond to the RCI-S² value (Berjanskii and Wishart 2008), the predicted secondary structure (red, helix; aqua, β-sheet), with the height of the bars reflecting the probability assigned by the SS-ANN secondary structure prediction. The bottom right panel depicts the χ¹ rotamer prediction (red oval: g−; green: g+; yellow: t), with the height of the ovals corresponding to the probability assigned by the χ¹ rotamer prediction. Note that in the top right panel, only two predictions (D52, G53) are deemed ambiguous. Both of these exhibit chemical exchange broadening in the NMR spectra, and adopt different type turns in different X-ray structures (type I in 1UBQ; type II in 3ONS) (Vijay-Kumar et al. 1987; Huang et al. 2012).

Fig. 5

TALOS-N χ¹ rotamer prediction for ubiquitin. The TALOS-N predicted χ¹ rotamer is displayed as a solid blue triangle. The χ¹ rotamers observed in four high resolution X-ray structure of ubiquitin, 1UBQ (1.8 Å, black circle), 1YJ1 (1.3 Å, red circle), 3A9J (1.18 Å, green circle) and 4HK2 (1.18Å, blue circle) are also shown in the top panel. The lower panel shows solvent accessibility, reported as the ACC parameter by the DSSP program (Kabsch and Sander 1983) (for ACC definition see http://swift.cmbi.ru.nl/gv/dssp/), as a function of residue number and averaged over the four X-ray structures.

Fig. 6

Summary of TALOS-N χ¹ rotamer prediction results for different residue types over the 34-protein validation set. The fraction of “Predictable” χ¹ rotamers (using a p_cutoff = 0.6 threshold) are marked as gray bars. Green bars depict the fraction of these rotamers whose prediction is consistent with the X-ray reference structure.